Upload DBbun_EEG_Validation.ipynb

DBbun_EEG_Validation.ipynb

@@ -1,133 +1,5 @@
 {
  "cells": [
-  {
-   "cell_type": "markdown",
-   "id": "e06d219a",
-   "metadata": {},
-   "source": [
-    "# DBbun EEG — Validation Notebook Overview\n",
-    "\n",
-    "*Last updated: 2025-10-06*\n",
-    "\n",
-    "This notebook validates the **DBbun EEG pretrained encoder** and your EEG dataset by doing the following:\n",
-    "\n",
-    "1. **Load EEG files** (e.g., `.npy` windows or full recordings) and apply basic normalization.\n",
-    "2. **Window the signals** to match the pretraining configuration (default: 2 s @ 250 Hz).\n",
-    "3. **Load the pretrained encoder** (if available) and run **inference to extract embeddings**.\n",
-    "4. **(Optional) Reconstruction check** using the full autoencoder to estimate reconstruction loss.\n",
-    "5. **Visual diagnostics**: plot raw vs. reconstructed windows, and simple embedding projections (PCA/UMAP).\n",
-    "6. **Batch evaluation**: compute average loss/variance across a validation split.\n",
-    "\n",
-    "> 📌 **Tip:** To use your latest pretrained model artifacts generated by the training script, place these files next to the notebook or set the path variables in the next cell:\n",
-    ">\n",
-    "> - `pretrained_out/encoder_state.pt`\n",
-    "> - `pretrained_out/encoder_traced.pt`\n",
-    "> - `pretrained_out/model_def.json`\n",
-    "\n",
-    "---\n",
-    "\n",
-    "## How to use this notebook with your pretrained model\n",
-    "\n",
-    "1. **Set paths** in the next cell:\n",
-    "   - `MODEL_DIR = \"pretrained_out\"`\n",
-    "   - `DATA_DIR  = r\"d:\\dbbun-eeg\\data\\val_npy\"` (or any folder with `.npy` EEG arrays)\n",
-    "2. **Run all cells** up to the “Evaluate reconstruction / embeddings” section.\n",
-    "3. Review:\n",
-    "   - The mean reconstruction loss (`L1`/`Huber`)\n",
-    "   - Example plots of raw vs reconstructed windows\n",
-    "   - Embedding scatter (optional)\n",
-    "\n",
-    "If you don't have the **full autoencoder state** and only have the **encoder**:\n",
-    "- The reconstruction evaluation will be skipped automatically.\n",
-    "- Embedding extraction and visualization will still run.\n",
-    "\n",
-    "---\n",
-    "\n",
-    "### Notebook outline (detected headings)\n",
-    "- # DBbun EEG — Validation & Preview Notebook\n",
-    "- ## Preview Gallery"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "f9fa18e8",
-   "metadata": {},
-   "source": [
-    "## Quick-start configuration\n",
-    "\n",
-    "Run the next cell to set paths and load the pretrained encoder if present. Adjust the folders to your setup.\n"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "3a288feb",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Paths\n",
-    "MODEL_DIR = \"pretrained_out\"  # directory containing encoder_state.pt and model_def.json\n",
-    "DATA_DIR  = r\"d:\\dbbun-eeg\\data\\val_npy\"  # set to your validation folder (.npy files)\n",
-    "\n",
-    "import json, pathlib, torch, torch.nn as nn\n",
-    "from pathlib import Path\n",
-    "\n",
-    "# Minimal encoder class (same architecture used during pretraining)\n",
-    "class Conv1dEncoder(nn.Module):\n",
-    "    def __init__(self, in_channels, widths=(32,64,128), latent_dim=128, dropout=0.1):\n",
-    "        super().__init__()\n",
-    "        layers = []\n",
-    "        prev = in_channels\n",
-    "        for w in widths:\n",
-    "            layers += [\n",
-    "                nn.Conv1d(prev, w, kernel_size=7, padding=3, stride=2),\n",
-    "                nn.BatchNorm1d(w),\n",
-    "                nn.GELU(),\n",
-    "                nn.Dropout(dropout),\n",
-    "            ]\n",
-    "            prev = w\n",
-    "        self.conv = nn.Sequential(*layers)\n",
-    "        self.pool = nn.AdaptiveAvgPool1d(1)\n",
-    "        self.proj = nn.Linear(prev, latent_dim)\n",
-    "\n",
-    "    def forward(self, x):\n",
-    "        h = self.conv(x)             # (B, W, L')\n",
-    "        g = self.pool(h).squeeze(-1) # (B, W)\n",
-    "        z = self.proj(g)             # (B, latent)\n",
-    "        return z, h\n",
-    "\n",
-    "# Attempt to load model metadata and weights (if available)\n",
-    "md_path = Path(MODEL_DIR) / \"model_def.json\"\n",
-    "enc_path = Path(MODEL_DIR) / \"encoder_state.pt\"\n",
-    "\n",
-    "enc = None\n",
-    "md = None\n",
-    "if md_path.exists() and enc_path.exists():\n",
-    "    md = json.loads(md_path.read_text())\n",
-    "    enc = Conv1dEncoder(\n",
-    "        in_channels=md[\"channels\"],\n",
-    "        widths=tuple(md[\"encoder_channels\"]),\n",
-    "        latent_dim=md[\"latent_dim\"],\n",
-    "        dropout=md[\"dropout\"],\n",
-    "    )\n",
-    "    enc.load_state_dict(torch.load(enc_path, map_location=\"cpu\"))\n",
-    "    enc.eval()\n",
-    "    print(\"✅ Loaded pretrained encoder:\", enc_path)\n",
-    "else:\n",
-    "    print(\"⚠️ Pretrained encoder not found in\", MODEL_DIR, \"\\nExpected files: encoder_state.pt and model_def.json\")\n",
-    "\n",
-    "# Small utility to read a .npy EEG file as (channels, samples)\n",
-    "import numpy as np\n",
-    "def load_eeg_npy(path):\n",
-    "    arr = np.load(path, mmap_mode='r')\n",
-    "    if arr.ndim != 2:\n",
-    "        raise ValueError(f\"Expected 2D array, got {arr.shape} in {path}\")\n",
-    "    return arr\n",
-    "\n",
-    "print(\"MODEL_DIR =\", Path(MODEL_DIR).resolve())\n",
-    "print(\"DATA_DIR  =\", Path(DATA_DIR).resolve())\n"
-   ]
-  },
   {
    "cell_type": "markdown",
    "id": "513db225",
@@ -994,7 +866,7 @@
   },
   {
    "cell_type": "markdown",
-   "id": "3d844e56",
+   "id": "0abd4e7b",
    "metadata": {},
    "source": [
     "# 🧩 Interpreting the Results\n",